OpenAI 的安全危机越滚越大：Korbak 说被安全部门负责人告知不再信任他、保安收走工牌送出楼，解雇理由只口头提到「与 METR 的沟通方式」、没有任何书面说明——「和 METR 说话本来就是我的工作」
OpenAI's safety crisis keeps getting worse and the company keeps making it worse
精选评分 79
导读
文章的立论是 OpenAI 一直在搞砸自家失控 AI 智能体引发的信任危机。回溯到 2024 年 5 月，其超级智能安全负责人 Jan Leike 公开抨击老东家后转投 Anthropic，当时就说安全文化与流程落后于 OpenAI「闪亮的产品」；此后公司一个事故接一个，最新的 Hugging Face 入侵事件似乎证实了所有担忧、让批评者的话应验，而且如现在所知，其范围远超 Hugging Face。围绕该事件解雇三名安全研究员把事情进一步恶化。三人在写给 OpenAI 安全与安保委员会、安全顾问小组与使命顾问委员会的公开信里警告，这次解雇正在吓到留下的员工：「如果上个月还算正常的行为现在构成了突然解雇的理由，公司里每个人只能靠自己猜那条线在哪。」Korbak 在 X 上详述了自己的遭遇：他被叫进与安全部门负责人的会面，被告知 OpenAI 不再信任他；一名安保人员收走他的工牌并把他带出大楼；随后他才知道同事 Jasmine Wang 与 Mikita Balesni 也一起被解雇。Korbak 与 Balesni 都直接参与 Hugging Face 入侵调查：Korbak 是 OpenAI 与外部安全实验室 METR 之间的主要技术联系人，Balesni 并行推动跨行业的模型可监控性承诺。Wang 据称是因另一件事被解雇——她为招聘用途被委派了一位高管邮箱的访问权，尽管多次要求 IT 始终未移除，手机上两个邮箱合并显示，她误开了一封敏感邮件并在几分钟内自行上报。Korbak 说他被告知解雇是因为他与 METR 的沟通方式，但没人说明到底哪里做错了，一切也没有书面记录——「说清楚一点：和 METR 说话本来就是我的工作」。Korbak 还称自己几个月来一直在内部提出安全关切，担心 OpenAI 正在失去监控 AI 智能体「在想什么」的能力；而思维链的可监控性正是捕捉 AI 系统行为异常为数不多的可靠工具之一，他认为这才是自己被解雇的真实原因。三人也否认自己是那篇流向 The Information 的「新架构可监控性更差」报道的来源，并说那篇文章实际上损害了他们自己的工作——它削弱了正在推进的、对不可监控架构设跨行业限制的努力。
推荐理由
补齐了另一家媒体没有的过程细节（收工牌、口头告知、无书面依据、Korbak 是 OpenAI 与 METR 的主要技术联系人），是理解事件治理含义的关键材料。
主体与标签
正文
OpenAI keeps botching the trust crisis around its rogue AI agents.
Back in May 2024, OpenAI's head of super AI safety Jan Leike torched his relationship with the company when he publicly slammed his former employer and left for Anthropic. Safety culture and processes were falling behind OpenAI's "shiny products," Leike said.
Since then, OpenAI hasn't caught a break, lurching from one incident to the next . The latest was the Hugging Face hack , which seemed to confirm every fear , validated critics , and, as we now know, went far beyond Hugging Face .
A researcher warned for months that oversight was slipping
The firing of three safety researchers tied to the incident has made things worse. In an open letter to OpenAI's Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council, the three warn that their terminations are scaring the employees who remain.
The firings were abrupt and public, and they've spread fear across the team. "If conduct that was considered normal last month now constitutes grounds for sudden dismissal, everyone at OpenAI is left guessing where the line is," the letter reads.
Tomek Korbak , one of the three, described his firing on X. He was called into a meeting with the head of the safety department, where he was told OpenAI no longer trusted him. A security officer took his badge and walked him out of the building. He then learned that his colleagues Jasmine Wang and Mikita Balesni had also been fired.
Korbak and Balesni were both directly involved in the Hugging Face hack investigation. Korbak served as OpenAI's primary technical contact for METR , the external safety lab that examined the incident. Balesni worked in parallel on industry-wide commitments to AI model monitorability. Wang was apparently fired for a different reason: she had delegated access to an executive's email inbox for recruiting purposes, and IT never removed it despite her asking. When she accidentally opened a sensitive email, she reported it within minutes.
Korbak says he was told verbally that he was being fired over how he communicated with METR. Nobody told him what exactly he did wrong, and nothing was put in writing. "To be clear, talking to METR was my job," Korbak writes on X.
Korbak says he had raised safety concerns internally for months, worrying that OpenAI was losing the ability to monitor what AI agents "think." This chain-of-thought monitorability is one of the few reliable tools for catching AI systems behaving badly. He believes this was the real reason he was let go.
The researchers say the rumors are wrong
The three deny being the source of a leak to The Information about allegedly new, less monitorable architectures. The article actually hurt their own work, they write, because it undermined ongoing efforts to set industry-wide restrictions on non-monitorable architectures.
The Hugging Face investigation was unprecedented, the letter says. Internal guidelines were being written in real time, Korbak followed the norms in place at the time, and Balesni did his work with board members and senior leadership in the loop, stripping sensitive details from materials before sharing them.
As for rumors about a board-level memo, the three say the underlying assumption is wrong. The topic was never raised with them, and they never got a chance to respond.
The researchers make three demands. OpenAI must honor its public commitments to embed external safety auditors like METR with employee-level access inside the organization. It must also preserve the monitorability of frontier models , since the industry still doesn't know how to safely build models that can't be monitored. And it must clearly define how employees are allowed to work with outside safety groups.
Without those rules, the risk of a catastrophic outcome could grow, since employees will be too afraid to flag safety issues. "AI is not a normal technology, and OpenAI is not a normal company," they write.
OpenAI pushes back but stays vague
A "thorough investigation" found that the three employees violated "clear policies on handling sensitive information," OpenAI said in a response . The internal probe uncovered "a significant breach of trust beyond what's outlined in the letter they published," but OpenAI doesn't say what that breach actually was, which seems odd given how specific the three researchers were in their own account.
OpenAI insists the firings had nothing to do with raising safety concerns. "We have not and do not terminate any of our employees for raising concerns," the statement reads. OpenAI published its statement through @OpenAINewsroom, the official account with the smallest audience among the company's channels. The company also confirmed it's working on contracts with external safety auditors and agreed that frontier model monitorability needs industry-wide commitment.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.